A variety of changes from Michael Hudson to get the compiler working
with 2.3. The primary change is the handling of SET_LINENO:
# The set_lineno() function and the explicit emit() calls for
# SET_LINENO below are only used to generate the line number table.
# As of Python 2.3, the interpreter does not have a SET_LINENO
# instruction. pyassem treats SET_LINENO opcodes as a special case.
A few other small changes:
- Remove unused code from pycodegen and pyassem.
- Fix error handling in parsermodule. When PyParser_SimplerParseString()
fails, it sets an exception with detailed info. The parsermodule
was clobbering that exception and replacing it was a generic
"could not parse string" exception. Keep the original exception.
There are now no known cases where the compiler package computes a
stack depth lower than the one computed by the builtin compiler. (To
achieve this state, we had to fix bugs in both compilers :-).
The chief change is to do the depth calculations with respect to basic
blocks. The stack effect of block is calculated. Then the flow graph
is traversed using breadth-first search to find the max weight path
through the graph.
Had to fix the StackDepthTracker to calculate the right info for
several opcodes: LOAD_ATTR, CALL_FUNCTION (and friends), MAKE_CLOSURE,
and DUP_TOPX.
XXX Still need to handle free variables in MAKE_CLOSURE.
XXX There are still a lot of places where the computed stack depth is
larger than for the builtin compiler. These won't cause the
interpreter to overflow the frame, but they waste space.
Remove the option to have nested scopes or old LGB scopes. This has a
large impact on the code base, by removing the need for two variants
of each CodeGenerator.
Add a get_module() method to CodeGenerator objects, used to get the
future features for the current module.
Set CO_GENERATOR, CO_GENERATOR_ALLOWED, and CO_FUTURE_DIVISION flags
as appropriate.
Attempt to fix the value of nlocals in newCodeObject(), assuming that
nlocals is 0 if CO_NEWLOCALS is not defined.
Fix list comp code generation -- emit GET_ITER instead of Const(0)
after the list.
Add CO_GENERATOR flag to generators.
Get CO_xxx flags from the new module
As the doc string for _lookupName() explains:
This routine uses a list instead of a dictionary, because a
dictionary can't store two different keys if the keys have the
same value but different types, e.g. 2 and 2L. The compiler
must treat these two separately, so it does an explicit type
comparison before comparing the values.
Use a dictionary instead of a list to map objects to their offsets in
a const/name tuple of a code object.
XXX The conversion is perhaps incomplete, in that we shouldn't have to
do the list2dict to start.
varnames should list all the local variables (with arguments first).
The XXX_NAME ops typically occur at the module level and assignment
ops should create locals.
Armin Rigo pointed out that the way the line-# table got built didn't work
for lines generating more than 255 bytes of bytecode. Fixed as he
suggested, plus corresponding changes to pyassem.py, plus added some
long overdue docs about this subtle table to compile.c.
Bugfix candidate (line numbers may be off in tracebacks under -O).
Fix annoying bugs in flow graph layout code. In some cases the
implicit control transfers weren't honored. In other cases,
JUMP_FORWARD instructions jumped backwards.
Remove unused arg from nextBlock().
pycodegen.py
Add optional force kwarg to set_lineno() that will emit a
SET_LINENO even if it is the same as the previous lineno.
Use explicit LOAD_FAST and STORE_FAST to access list comp implicit
variables. (The symbol table doesn't know about them.)
embedded code objects (e.g. functions) rather than the generated code
object. This change means that the compiler generates code for
everything at the end, rather then generating code for each function
as it finds it. Implementation note: _convert_LOAD_CONST in
pyassem.py must be change to call getCode().
Other changes follow. Several changes creates extra edges between
basic blocks to reflect control flow for loops and exceptions. These
missing edges had gone unnoticed because they do not affect the
current compilation process.
pyassem.py:
Add _enable_debug() and _disable_debug() methods that print
instructions and blocks to stdout as they are generated.
Add edges between blocks for instructions like SETUP_LOOP,
FOR_LOOP, etc.
Add pruneNext to get rid of bogus edges remaining after
unconditional transfer ops (e.g. JUMP_FORWARD)
Change repr of Block to omit block length.
pycodegen.py:
Make sure a new block is started after FOR_LOOP, etc.
Change assert implementation to use RAISE_VARARGS 1 when there is
no user-specified failure output.
misc.py:
Implement __contains__ and copy for Set.
1.5.2. The compiler generates code for the version of the interpreter
it is run under.
ast.py:
Print and Printnl add dest attr for extended print
new node AugAssign for augmented assignments
new nodes ListComp, ListCompFor, and ListCompIf for list
comprehensions
pyassem.py:
add work around for string-Unicode comparison raising UnicodeError
on comparison of two objects in code object's const table
pycodegen.py:
define VERSION, the Python major version number
get magic number using imp.get_magic() instead of hard coding
implement list comprehensions, extended print, and augmented
assignment; augmented assignment uses Delegator classes (see
doc string)
fix import and tuple unpacking for 1.5.2
transformer.py:
various changes to support new 2.0 grammar and old 1.5 grammar
add debug_tree helper than converts and symbol and token numbers
to their names
- fix tab space issues (SF patch #101167 by Neil Schemenauer)
- fix co_flags for classes to include CO_NEWLOCALS (SF patch #101145 by Neil)
- fix for merger of UNPACK_LIST and UNPACK_TUPLE into UNPACK_SEQUENCE,
(SF patch #101168 by, well, Neil :)
- Adjust bytecode MAGIC to current bytecode.
TODO: teach compile.py about list comprehensions.
Attached is a set of diffs for the .py compiler that adds support
for the new extended call syntax.
compiler/ast.py:
CallFunc node gets 2 new children to support extended call syntax -
"star_args" (for "*args") and "dstar_args" (for "**args")
compiler/pyassem.py
It appear that self.lnotab is supposed to be responsible for
tracking line numbers, but self.firstlineno was still hanging
around. Removed self.firstlineno completely. NOTE - I didnt
actually test that the generated code has the correct line numbers!!
Stack depth tracking appeared a little broken - the checks never
made it beyond the "self.patterns" check - thus, the custom methods
were never called! Fixed this.
(XXX Jeremy notes: I think this code is still broken because it
doesn't track stack effects across block bounaries.)
Added support for the new extended call syntax opcodes for depth
calculations.
compiler/pycodegen.py
Added support for the new extended call syntax opcodes.
compiler/transformer.py
Added support for the new extended call syntax.
code generator uses flowgraph as intermediate representation. the old
rep uses a list with explicit "StackRefs" to indicate the target
of jumps.
pyassem converts flowgraph to bytecode, breaks up individual steps of
generating bytecode
- removed now (happily) unused second arg
- need to verify results of [].index are correct; for building consts,
need to have same value and same type, e.g. 2 not the same as 2L
(big surprise). new solution is a little less hackish.
Code gen adds a TupleArg instance in the argument slot. The tuple arg
includes a copy of the names that it is responsble for binding. The
PyAssembler uses this information to calculate the correct argcount.
all fix this wacky case: del (a, ((b,), c)), d
which is the same as: del a, b, c, d
(Can't wait for Guido to tell me why.)
solution uses findOp which walks a tree to find out whether it
contains OP_ASSIGN or OP_DELETE or ...
Jeremy Hylton
Neal Norwitz
Tim Peters
Thomas Wouters